International Journal of Computer Science Trends and Technology (IJCST) – Volume 6 Issue 6, Nov-Dec 2018

RESEARCH ARTICLE OPEN ACCESS

Feed Forward Back Propagation Neural Network for Detection of

Breast Cancer
Vijaylaxmi Kochari
Bharatesh College of Computer Applications
Belagavi, Karnataka
India

ABSTRACT
Breast cancer can be detected at a very early stage by using Breast screening method. First X-ray of each breast is taken, called
mammogram. Cancers which are difficult to identify by the doctor or felt by the women herself can be detected through these
mammograms. These mammograms can identify the small changes in breast tissue which may cause cancer. The help of
technology can improve the diagnosis accuracy. So in this work Feed Forward Back Propagation Neural Network (FFBPNN)
classifier is used to classify the mammogram as either benign or malignant. In this process first the mammogram is
preprocessed then the mammogram is segmented using K means clustering algorithm, features of segmented mammogram are
extracted, the models are built through training and then Feed Forward Back Propagation Neural Network (FFBPNN) classifier
is used to classify the mammogram as either benign or malignant.
Keywords :— Mammogram, Preprocess, Segmentation, FFBPNN

I. INTRODUCTION II. DATA SOURCE

Breast cancer is one of the main causes of death among
women and the most frequently diagnosed non-skin cancer in
women [1]. Breast cancer occurs when the cell tissues of the
breast become abnormal and uncontrollably divided. These
abnormal cells form large lump of tissues, which consequently
becomes a tumor [2]. If these disorders are detected at early
stage then they could be successfully treated. So it is very
much important detect the breast cancer signs as early as
possible through appropriate methods.
In most of the hospitals and clinics Mammography method is
used to detect the breast cancer at the earliest stage. It has
been proven effective to reduce mortality as much as by 30%
[6]. The main purpose of screening mammography is to
identify the cancerous tumor as early as possible and eliminate
it before the establishment of metastases. The early signs for
breast cancer are masses and microcalcification but the
abnormalities and normal breast tissues are often difficult to
be differentiated due to their subtle appearance and ambiguous
margin. Only about 3% of the essential information are
discovered during a mammogram where a part of suspicious
area is enclosed with vessels and normal tissues. This
circumstances may cause the radiologists hard to make out a
cancerous tumor. To read the mammograms much better and
assists the radiologists, computer-aided diagnosis has been
developed. Neural Network model can be used in computer-
aided diagnosis for mammography interpretation and biopsy
decision making.
Different types of mammograms are taken from standard
MIAS (Mammographic Image Analysis Society) database for
training as well as testing purpose. MIAS database of digital
mammograms is generated by UK research group. The
database contains 161 patients’ right and left breast images.
So totally the database contains 322 images. To make all the
images of same size i.e. 1024 x 1024 pixel the database has
been reduced to 200-micron pixel.
III. SYSTEM DESIGN
The various tasks which are carried out in this work are image
preprocessing, segmentation, feature extraction, building
models through training and then using classifier to classify
these images as benign or malignant.
A. Preprocessing
Operations on images at the low level of abstraction are
present in image preprocessing. Adaptive histogram
equalization is one of the preprocessing methods realized in
MATLAB. It is mainly used to improve quality of the image
data that suppresses unwanted distortion and it is also used to
enhance some image features which are very necessary for
further processing of the mammogram. Image information
content is not increased by image preprocessing, but the
redundancy in the image is used by the image preprocessing.
Corresponding to one real object neighboring pixels have the
same or similar brightness value. Distorted pixel can be
restored in image as an average value of neighboring pixels.

ISSN: 2347-8578 www.ijcstjournal.org Page 19

 International Journal of Computer Science Trends and Technology (IJCST) – Volume 6 Issue 6, Nov-Dec 2018
Fig. 1 Original Mammogram
Fig. 2 Mammogram after applying Adaptive Histogram Equalization
Fig. 1 shows the original mammogram selected from standard
MIAS database. After application of Adaptive Histogram
Equalization mammogram is improved. The improved
mammogram is displayed in the fig. 2. In fig. 1 the regions
which had low contrast are improved and shown in fig. 2.
B. Segmentation
Dividing an image into distinct regions where the pixels of
each region have similar characteristics is called segmentation.
The success of image analysis depends on an effective image
segmentation process. K-Means clustering algorithm is used
to device the segmentation of mammogram. The algorithm
works as follows,
K Means clustering algorithm:
Step 1: Number of clusters must be known previously to be K
Step 2: Select K number of cluster centers such that they are
farthest apart from each other
μi= some value , i=1,...,k
Step 3: Consider each pixel and assign it to the cluster which
is closest
ISSN: 2347-8578 www.ijcstjournal.org
ci= {j : d(xj, μi) ≤ d(xj, μl), l≠i, j=1,...,n}
Step 4: Recalculate cluster centers by finding mean of pixels
belonging to the same cluster
μi= 1/|ci| ∑j∈cixj, ∀i
|c| = number of elements in c
Step 5: Repeat step 3 and step 4 till shifting of cluster centers
are observed [3].
Fig. 3 Mammogram after applying K Means Clustering
Fig. 4 Segmented Mammogram
The K Means Clustering algorithm is enforced on the
enhanced image. Here the value of K is three. So total three
clusters are formed. The clustered image is shown in fig. 3.
The black cluster shows the area which is of not at all useful.
The gray cluster is also not so important. The brightest cluster
shows the area which is of very important. This brightest
cluster is selected and segmented. The segmented
mammogram is shown in fig. 4.
Image segmentation is a method used in image processing.
Mainly, segmentation is done on the raw image to identify
small, local, and bright spots. Research done by Kevin et al.
Page 20
 International Journal of Computer Science Trends and Technology (IJCST) – Volume 6 Issue 6, Nov-Dec 2018
has drawn significant attention on the segmentation process input values be x1, x2, x3 and output be y as represented in fig.
and neural network used [4]. 5.
The accuracy of the FFBPNN is tested by having a set of
labelled test images for determination of true positive (TP)
and false positive (FP) detection rates.

C. Feature Extraction

A discrete cosine transform (DCT) represents a limited

sequence of data points in the form of addition of cosine
functions fluctuating at various frequencies. DCT is same as
DFT and deals with Fourier related transformations, and it
only makes use of real numbers. After applying DCT on a
mammogram thirty five features are extracted for each
mammogram and stored in the database. These features are
used by the FFBPNN classifier.
D. Building models through training and
Classifier All these values x1, x2, x3 are fed as input to NN, through the
In this the classifier used is FFBPNN. Let us see how this NN
works and how it is trained and implemented.
Facts of Neural Networks (NN) are
 Various layers are present in the NN.
 There is no limit on the number of neurons in every
layer.
Structure of a NN is as explained below
 Inputs are applied at the first layer and it called input
layer.
 Outputs are produced at the last layer and it is called
as output layer.
 Between the input and output layer of NN, there can
be any number of hidden layers.
 In most neural network models, a neuron which is
present in one layer is connected to all neurons which
are present in the next layer.
Basic facts about training of NN
 Provide proper inputs and outputs to train the NN.
 The network learns from the inputs and outputs.
 After the training gets finished, then give proper
input to get the output.
NN are parallel data handling frameworks and consists of
neurons. All the neurons in NN are arranged in layer format
and are associated with each other through links. The brain
structure and nervous system of people and creatures are
strongly connected; this same system is created in artificial
NN also. Axons work is imitated by links in the artificial NN
and cells in human body are represented as neurons in
artificial NN [5]. Different types of neural networks are there.
In this project feed forward neural networks are used. If we
take a case of a multilayer Feed Forward Back Propagation
Neural Network (FFBPNN) where we are going to provide
three inputs and we are going obtain one output. Let these
Fig. 5 Representation of FFBPNN
connection links it distributes the information. There are
weights w'ij known as connection weights present in the
middle of input layer and hidden layer, these weights are
multiplied to the input values. The function fhcalled as
activation function gets the input from the hidden neurons.
This input is calculated by adding the signals from neurons
present at the input layer. The w''j connections which are
present in the middle of neurons of hidden layer and output
layer are weighted by resulting activations of neurons present
in the hidden layer and propagated to the neuron(s) present in
the output layer. The projection and addition on the function fo
is performed by the neuron present at the output layer. The
projected response is the output of the neuron present in the
output layer. The projected answer for the single neuron
present at the output layer is:
Where number of input neurons is represented by nv, the
number of hidden neurons are represented by nh. Bias of the
neurons present in the hidden layer are represented by and
bias of the neurons present in the output layer is represented
by . The transfer functions are horizontally shifted by
them. Random values are assigned to weights and biases
before calibration, these values can also be modified by
learning algorithms during training process. The output value
i.e. recognized y value is passed in the network through the
training process. If the predicted response is different from
that of experimental response then difference between them is
calculated and this difference is taken as error value. This
estimated error is propagated back and the parameters are
adjusted so that the error is minimized. Till the error value is
minimized, these two steps are repeated. These steps frame an
epoch. From this it is clear that the learning algorithm always
tries to minimize the error, and it will be always trying to get
an acceptable low level error value.

ISSN: 2347-8578 www.ijcstjournal.org Page 21

 International Journal of Computer Science Trends and Technology (IJCST) – Volume 6 Issue 6, Nov-Dec 2018

Implementation of FFBPNN

The extracted features after applying DCT on mammogram

are supplied as input to the feed-forward back propagation
neural network at the input layer. Random weight and bias
values are assigned and output is calculated by summation
unit and transfer function sends output to the next hidden layer.
For all the hidden layers this process is continued. At the
output layer there is only one neuron which can classify the
mammogram as benign or malignant. Then the output is
compared with actual output if there is a difference between
them it is calculated and treated as error. If any error is there it
is calculated and it is propagated back to the last hidden layer.
This NN is called Backward Propagation Neural Network, Fig. 6 Sequence diagram of the proposed system
because there is propagation of delta values or error values
backward. Depending on the delta value the weight and bias
values are changed and process is repeated till there is no error.
V. RESULT ANALYSIS
Once this training process is over, the next step is testing. In
testing process the mammogram which needs to be tested is We can find the accuracy of the classifier by analyzing the
preprocessed, segmented and then DCT is applied and results. Results can be analyzed by checking True Positive,
features of the segmented mammogram are extracted. These True Negative, False Positive and False Negative output
features are sent as input to the feed-forward back propagation values of the classifier. From the Table 1 it is clear that the
neural network and it will classify the mammogram as either proposed approach which is based on FFBPNN got a better
benign or malignant. results and classification rate of 96.66%.
To create FFBPNN MATLAB provides a function
Mammograms Tested TP TN FP FN
“newff” and “sim” simulate a Simulink model. Inbuilt
function “sim” compares the extracted features of the Normal 30 29 1 0 0
mammogram under test conditions to all the mammograms in
the database, if the features are matched it displays the Abnormal 30 29 1 0 0
appropriate result to the user.
Total 60 58 2 0 0

IV. SEQUENCE DIAGRAM

Table 1 Result analysis on different Mammograms

A sequence diagram is a representation of series of message

communications between the participating members. In fig. 6
the user first loads the mammogram which is then subjected to
preprocessing followed by segmentation. Features of the
segmented mammogram are extracted followed by training
which trains the system using FFBPNN. In the testing phase
the user follows the same procedure till feature extraction then
image undergoes testing and then the appropriate result is
displayed to the user. Then results are analyzed by checking
the training time and accuracy of the classifiers.
VI. CONCLUSION
The system takes mammogram, preprocesses it by applying
Adaptive Histogram Equalization. By using K Means
Clustering algorithm, enhanced image is segmented. Features
of segmented mammogram are extracted. These extracted
features are fed as input to the classifier Feed Forward Back
Propagation Neural Network. This classifies the mammogram
as benign or malignant. The proposed approach which is
based on FFBPNN got a better results and classification rate
of 96.66%. This will help the doctors to diagnose breast
cancer and help to save the lives of many women’s.
In future accuracy can be increased by using other
classification methods or by extracting other features.

REFERENCES

[1] L. Hadjiiski, B. Sahiner, M. A. Helvie et al., “Breast

masses: computer-aided diagnosis with serial

ISSN: 2347-8578 www.ijcstjournal.org Page 22

International Journal of Computer Science Trends and Technology (IJCST) – Volume 6 Issue 6, Nov-Dec 2018
mammograms,” Radiology, vol. 240, no. 2, pp. 343–
356, 2006.
[2] T. Balakumaran, I. L. A. Vennila, and C. G. Shankar,
“Detection of microcalcification in mammograms
using wavelet transform and fuzzy shell clustering,”
International Journal of Computer Science and
Information Technology, vol. 7, no. 1, pp. 121–125,
2010.
[3] http://www.onmyphd.com/?p=k-
means.clustering&ckattempt=1
[4] M. M. Mehdy , P. Y. Ng , E. F. Shair , N. I. Md
Saleh , and C. Gomes “Artificial Neural Networks
in Image Processing for Early Detection of Breast
Cancer”, Computational and Mathematical Methods
in Medicine Volume 2017, Article ID 2610628
[5] http://www.frank-dieterle.com/phd/2_7_1.html
[6] J. Dheeba, N. Albert Singh, and S. Tamil Selvi,
“Computer-aided detection of breast cancer on
mammograms: a swarm intelligence optimized
wavelet neural network approach,” Journal of
Biomedical Informatics, vol. 49, pp. 45–52, 2014.

ISSN: 2347-8578 www.ijcstjournal.org Page 23